Clinical evaluation on image quality of a deep learning-based denoising algorithm in 18F-FDG PET/CT studies

1532 Objectives: This study is to investigate the performance of a newly-developed deep learning-based reconstruction algorithm (HYPER DLR) and its impact on image quality as compared with standard ordered-subsets expectation maximization (OSEM). Methods: Fifty-two oncological patients were retrospectively enrolled who underwent a FDG PET/CT scan on a digital PET/CT (uMI 780, United Imaging Healthcare). The acquired list-mode PET rawdata were then cut into 180s, 135, 90s, and 60s group with two reconstruction algorithms, OSEM and HYPER DLR. HYPER DLR was based on a variant of U-Net structure, trained with over 300 cases and validated with other 80 cases. The overall image quality was first assessed by two experienced nuclear medicine physicians with a 5-point Likert scale. And liver SUVmean, SUVmax, SD and lesion SUVmax were measured by manually drawing a region of interest (ROI) in all the series of PET images. Subsequently, liver signal-to-noise ratio (SNR) and lesion target-to-background ratio (TBR) were obtained. Equivalence tests were used to quantify the agreement on SUV between the groups while other metrics were compared with paired-samples t test. The comparison was performed between OSEM and HYPER DLR with the same acquisition time and OSEM_180 with all DLR groups. Results: The overall inter-reader agreement of the image quality indicated a substantial agreement with a kappa of 0.705. The DLR group showed a higher score when comparing to the OSEM group with the same acquisition time. Compared to OSEM_180 group, the DLR_135 group showed no significant difference while other DLR groups showed a significant difference on the image quality score (p<0.001). The equivalence tests showed equivalent performance of SUV in livers and lesions between groups. The liver SD in DLR groups showed a significant decrease than OSEM groups, while the SNR showed a significant increase as shown in Fig. 1A (all p<0.001). It indicated the capability of the HYPER DLR algorithm in the noise reduction in the liver image as shown in Fig. 1B. The DLR group showed a higher TBR than one of OSEM group with the same acquisition time (p≤0.001) indicating the increased lesion conspicuousness. Compared to the OSEM_180 group, the DLR_135 group showed no significant difference (p=0.713) while other DLR groups were significantly lower on the lesion TBR as shown in Fig. 1C (p<0.001). Conclusions: The developed deep-learning algorithm showed relevant improvement in image quality compared with OSEM reconstruction in PET/CT oncological studies. HYPER DLR reconstruction yielded improved results in the liver SNR and lesion TBR which indicated a better image quality and lesion conspicuousness. Moreover, when comparing the results with the reference group, it offers potential applications in PET imaging with shorter acquisition time or reduced FDG injection dose.